For more than a decade, progress in artificial intelligence has followed a similar pattern: build larger models, train them on larger datasets and reap consistent gains. This ‘age of scaling’ has reshaped industries, accelerated digital transformation and set new expectations for automation and intelligence.
But as one of the field’s most influential figures recently argued, that era appears to be ending. In an insightful interview, machine learning expert Ilya Sutskever described a shift underway at the heart of AI research from raw scale to deeper understanding. Despite increasingly promising benchmark performances, he says, today’s systems still sometimes behave inconsistently in complex, real-world environments. Scaling pushed performance upwards through size, but did not address the gaps in judgement, intuition and generalisation. The next era aims to close those gaps, making systems not just more powerful but fundamentally more reliable.
This issue is also causing a tangible tension for businesses, with many leaders feeling the urgency to adopt AI, yet struggling to align ambition with capability. In our recent research, 92% of fintech and financial services leaders say they feel prepared to embed AI, but only 36% have a fully funded, enterprise-wide strategy to guide that shift.
Why scaling alone is no longer enough
The last five years of AI development produced extraordinary advances but also exposed a deeper constraint. Models can write code, summarise documents and simulate human conversation, yet can show brittleness when circumstances shift. Sutskever likens this to students who have memorised every trick to pass an exam but lack the intuitive flexibility of those who truly ‘have it’.
For organisations, this is more than a conceptual point. It surfaces in every sector, with challenges taking different forms:
- In financial services, a model that performs well in testing may behave unpredictably within tight regulatory constraints.
- In healthcare, inconsistency poses risks where outcomes depend on precision and reliability.
- In retail and travel, even small fluctuations in performance can undermine customer trust.
- In government, brittleness amplifies complexity, affecting public services that cannot afford error.
These concerns were highlighted in our research, with barriers to AI adoption including data quality and integration challenges, as well as concerns over trust, transparency and explainability.
Yet, leaders still feel the need to adopt this technology and build institutional knowledge, or they risk falling behind competitors. To combat this tension, leaders must take an approach of understanding and strategic implementation.
A new era: from scaling to understanding
According to Sutskever, AI is entering a new chapter in which breakthroughs will come from conceptual progress rather than expanding model size. This shift could focus on several key areas:
- Better generalisation
Humans can master domains they have never encountered before but today’s AI models struggle with this kind of adaptive reasoning. Closing that gap will require new architectures and learning mechanisms, rather than more compute power. - Continuous self-correction
One of Sutskever’s most interesting observations is that people evaluate their trajectory moment by moment. We sense when something feels off long before a failure occurs. Models, by contrast, often only receive feedback once an entire task is complete. Introducing forms of internal grounding – a ‘value function’ equivalent – could dramatically improve stability and reliability. - Moving towards agentic systems
Agentic systems can plan, reason and act across contexts. Ensuring the underlying intelligence is robust will be crucial in ensuring these systems become reliable collaborators.
What this shift means for leaders
The transition from scaling to understanding reshapes the fundamentals of how organisations use and prepare for AI.
Competitive advantage will come from adaptability, not access
Access to large models is becoming widely available. The differentiator is how quickly an organisation can interpret new capabilities, experiment safely and integrate them into processes. Institutional knowledge becomes the strategic asset.
Architectures must become modular and replaceable
As the underlying recipe for AI evolves, organisations will need systems that accommodate different models, reasoning layers and evaluation mechanisms. Legacy complexity already slows innovation, and leaders must prioritise laying the foundations with architecture that can flex as the frontier shifts.
Decision-making frameworks must evolve with the technology
Traditional project-based approval cycles struggle with technologies that change monthly. Many businesses are already looking to evolve decision-making, with 12% of leaders having already adopted a fully AI-driven approach.
To do so, organisations will need dynamic governance that looks at:
- where autonomy is appropriate
- how escalation and oversight are structured
- how to respond when model behaviour shifts
Human roles will continue to evolve
As systems take on more, human roles will focus on supervision, context, judgement and exception handling. Whether in financial services, healthcare, retail or the public sector, success will rely on people who can understand how AI fits into broader organisational goals.
Leading in the age of understanding
The scaling era delivered remarkable progress, but we do not yet have fully intuitive, reliable and context-aware systems. The next decade will be shaped by breakthroughs in understanding, from better generalisation and real-time self-correction to increasingly agentic behaviour.
This shift creates both challenge and opportunity. The organisations that prepare now by partnering with leading experts, building adaptable architectures and fostering a culture of continual learning will be positioned to navigate whatever breakthroughs come next.
See how your organisation compares with others making the shift by downloading our report.